Personalised multimodal models
Explore physiological time series alongside movement, task events, and participant reports. Study individual adaptation and evaluate how models behave across people, sessions, and settings.
Research vertical 02 / ManasAICutting-edge
Cutting-edge
AI applications.
Intelligence that understands context deserves equally thoughtful evaluation.
We work at the intersection of multimodal learning, language models, and personalised inference. Our focus is on systems that combine useful predictions with a clear account of their evidence, context, and limitations.
MANASAI / RESEARCH 02CONCEPTUAL VIEW
The research question
What makes an AI model useful for a particular person?
A person’s baseline, the task they are performing, and the way a question is framed all matter. We investigate how AI can combine these sources of context, and how its behaviour changes when they do. The goal is to develop applications that researchers can scrutinise, adapt, and evaluate.
What we investigateThree connected
Three connected
research directions.
Questions guiding this vertical and its contribution to intelligence for inner wellbeing.
Language and knowledge systems
Develop retrieval and reasoning approaches for domain-specific and Indic knowledge. Source traceability, cultural context, and careful evaluation guide this work on reflective and research-facing applications.
Evaluation beyond accuracy
Investigate probability consistency, contextual framing, and order effects. A model that answers well in one prompt may behave differently under a changed context; those differences belong in the evaluation.
How we approach it
From a useful question to a tested application.
Frame
Define the use case and what a useful answer would mean.
Ground
Connect relevant signals, sources, and study context.
Evaluate
Test accuracy, consistency, uncertainty, and variation.
Personalise
Study adaptation to individuals without losing sight of limitations.
Research that informs the directionIdeas with
Ideas with
a published trail.
Selected work by our scientific lead and collaborators. Each paper is credited to its authors; publication status is shown alongside the source.
Emergent non-classical probabilistic structure in large language models under contextual modulations
An investigation of context-sensitive probability judgements in six open-weight language models. The reported departures from classical probability vary by model; they do not establish a universal quantum signature.
Research implication: evaluate consistency, framing, and question-order effects alongside answer accuracy.
Read published articleBeyond alignment: representational ethics and the governance of constructed worlds
A framework for examining how AI shapes meaning and experience. It extends ethical analysis from observable behaviour to the representations through which people understand and act in the world.
Design implication: assess how feedback frames experience and affects agency, as well as whether a model’s output is accurate.
Read author manuscriptThese works inform the research agenda. They do not constitute validation of SakshiSense hardware or wellness outcomes.
Connection to inner wellbeing
Wellness AI with context built in.
For stress and attention research, a prediction only becomes meaningful in the setting of the study. This vertical develops the inference and evaluation methods that can connect SakshiSense data to personalised models and, over time, carefully designed feedback.
Explore the wellness platform ↗Research with ManasAI
Discuss an AI collaboration ↗Building an AI application?
Bring a use case, a dataset, or an evaluation challenge. Let’s define the research together.